Abstract
We first extend the fuzzy regression analysis formulated with symmetric triangular fuzzy number coefficients to the case of non-symmetric triangular and trapezoidal fuzzy numbers. We show that some difficulties in the fuzzy regression analysis can be remedied by this extension. Next we propose an approach to non-linear fuzzy regression using fuzzy neural networks with fuzzy number connection weights. Leaning methods are proposed for the following three kinds of training data : real number data, real number data with membership values, and fuzzy data with fuzzy number outputs. Last we apply the proposed approach to the determination problem of type 2 membership functions. A membership value is given as a fuzzy number such as "small" and "large" in a type 2 membership function.
